Definition
An explore agent is an agent role optimized for mapping and scoping: it reads, searches, traces dependencies, and produces a plan. It should usually not make code changes.
Think: “survey crew,” not “construction crew.”
Core outputs
A good explore pass produces:
- A short problem statement in plain language
- Pointers into the codebase (files, functions, call chains)
- Known constraints (versions, policies, invariants)
- A proposed sequence of small changes
- Risks and “unknowns”
- A build receipt showing what it inspected (commands run, files read)
When to use
- Large or unfamiliar codebases
- Issues that smell like “it’s not where you think”
- Changes with hidden blast radius (security, auth, build tooling)
- Any time you’re seeing context starvation in task execution
How it fits in a flow
Explore agents are most useful as the first stage in a gated multi-agent flow:
- Explore → plan + boundaries
- Implement → task agent executes
- Verify → gates + adversarial checks
Failure modes
- “Exploration” turns into unbounded ideation (no plan, no stop condition).
- The explore agent starts editing code, creating hidden state and hard-to-review diffs.
- The output is narrative-only (no pointers, no receipts).
Practical rule
If you can’t point to where in the repo the answer lives, you’re not done exploring.